Multi-dimensional on-line analysis system for sphericity of gas atomized powder and method thereof

Abstract

The present application relates to the technical field of quality detection, and particularly relates to an aerosolized powder sphericity multidimensional online analysis system and method thereof, which comprises a dynamic double-mode imaging module, a transverse wind field auxiliary detection unit and a multidimensional feature fusion analysis module, the transverse wind field auxiliary detection unit is used for applying controllable transverse wind force through airflow nozzles orthogonal to the falling direction of the powder, and combining with a laser displacement sensor to measure the deflection trajectory of aerosolized powder spherical particles; and the present application further comprises an online calibration module, which is used for spraying standardized spherical particles to the powder flow at a preset period; compared with the traditional density detection which relies on destructive sampling and offline measurement, the present application has low efficiency and cannot cover all the particles on the production line; through the orthogonal wind field trajectory inversion technology, the present application synchronously outputs the density value and internal defect mark through non-contact dynamic measurement, realizes 100% nondestructive online full detection on the production line, and breaks through the sampling limitation and time bottleneck of the traditional means.

Description

Technical Field

[0001] This invention relates to the field of quality testing technology, and in particular to a multi-dimensional online analysis system and method for the sphericity of atomized powder. Background Technology

[0002] Density and internal defect detection of atomized powders are crucial for ensuring the quality of additive manufacturing and powder metallurgy products. Currently, the industry mainly relies on offline sampling combined with the specific gravity bottle method or helium porosimeter to measure density. This requires destructive sample preparation and takes more than 30 minutes per test, achieving a sampling rate of only 0.1%-1%, resulting in a large number of defective particles flowing into downstream processes. For hidden defects such as internal porosity and component segregation, traditional two-dimensional imaging technology is limited by surface morphology analysis. Although X-ray computed tomography (CT) can reveal internal structures, its detection speed is slow (minutes per particle) and the equipment cost is high, failing to meet the needs of continuous online analysis on production lines.

[0003] Existing technologies suffer from three major bottlenecks: First, destructive density testing wastes raw materials and cannot achieve full inspection, with sampling statistical biases leading to the risk of batch quality loss of control. Second, surface imaging technology has a detection rate of less than 60% for internal defects and cannot quantify porosity distribution and compositional segregation. Third, offline detection data is lagging, with process control response delays exceeding one hour, resulting in persistently high scrap rates. These shortcomings severely restrict the production yield and cost control of high-value-added metal powders (such as aerospace-grade titanium alloys and high-temperature alloys).

[0004] This invention addresses the aforementioned problems through orthogonal wind field trajectory inversion and multiphysics fusion perspective technology: by using non-contact lateral wind loading and laser displacement tracking, it inverts the density of individual particles in real time and simultaneously marks internal defects, achieving 100% online non-destructive inspection; by fusing visible light, infrared thermography, and dynamic data, combined with a physically constrained 3D reconstruction algorithm, it penetrates the particle surface, making industrial-grade visualization of internal porosity and segregation possible. This solution improves detection efficiency by 100 times, achieves a hidden defect detection rate exceeding 98%, and simultaneously drives millisecond-level closed-loop optimization of the atomization process, providing a highly efficient quality control tool for the gas atomization powder production field. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes a multi-dimensional online analysis system and method for the sphericity of gas-atomized powder.

[0006] The technical solution of this invention is: a multi-dimensional online analysis system for the sphericity of gas-atomized powder, comprising:

[0007] The dynamic dual-mode imaging module includes a visible light industrial camera and an infrared thermal imager, used to simultaneously acquire surface morphology images and temperature field distribution of atomized powder spheres;

[0008] The lateral wind field auxiliary detection unit is used to apply controllable lateral wind force through airflow nozzles orthogonal to the direction of powder fall, and combine it with a laser displacement sensor to measure the deflection trajectory of atomized powder balls.

[0009] The multi-dimensional feature fusion analysis module is used to perform multi-dimensional feature analysis on gas-atomized powder particles by fusing morphology, temperature and kinetic data based on AI algorithms.

[0010] As a preferred embodiment, the dynamic dual-mode imaging module is configured as follows:

[0011] A11: Employs a high-speed adaptive imaging method to dynamically adjust the frame rate of the visible light industrial camera based on the velocity of the atomized powder spheres.

[0012] A12: It adopts a multi-point sampling array, and the imaging points composed of the visible light camera and the infrared thermal imager are set to four or more groups, and the multiple imaging points are distributed along the transmission path of the atomized powder ball particles.

[0013] A13: Optical flow motion correction is used to eliminate the impact of motion blur on image analysis.

[0014] As a preferred embodiment, the dynamic dual-mode imaging module dynamically adjusts the frame rate of the visible light industrial camera according to the velocity of the atomized powder spheres, and the specific process is as follows:

[0015] S11: Real-time acquisition of particle velocity in powder flow via laser velocity sensor;

[0016] S12: Calculate the target frame rate based on the movement speed of the atomized powder spheres using the frame rate calculation formula, where the frame rate calculation formula is:

[0017] ;

[0018] in, The velocity of the atomized powder particles. The minimum target particle size for atomized powder spheres. The average particle size of the atomized powder spheres. This is the fluctuation compensation coefficient. Target frame rate;

[0019] S13: Dynamically adjust the frame rate of the visible light industrial camera and the infrared thermal imager to the target frame rate, and synchronously trigger the pulse flash exposure time, wherein the pulse flash exposure time... The calculation formula is:

[0020] ;

[0021] S14: Determine if a single particle has been continuously captured for at least 3 frames. If not, trigger secondary frame rate compensation.

[0022] ;

[0023] in, The frame rate after secondary compensation. The number of missing particles in the expected number of capture frames is counted in real time by the image matching algorithm.

[0024] As a preferred option, when using optical flow motion correction to eliminate the impact of motion blur on image analysis, the specific process of the dynamic dual-mode imaging module is as follows:

[0025] S21: Extract the SIFT feature point set of the particle surface from the continuously captured frame sequence;

[0026] S22: Based on the improved Lucas-Kanade optical flow method, the displacement vector field of characteristic points is calculated. The principle formula is as follows:

[0027] ;

[0028] in, For the displacement vector field to be solved, For granular regions, To smooth the constraint coefficients, For image spatial gradient, The image's temporal partial derivative;

[0029] S23: Decompose the displacement field into rigid body translational and rotational components:

[0030] , ;

[0031] in, For rigid body translation components, This represents the number of effective feature points on the particle surface. The horizontal coordinates of the feature point in the image coordinate system. The vertical coordinates of the feature point in the image coordinate system. For point Displacement vector at that point For rotational components, It is a median function. This is the perpendicular component of the displacement vector. The horizontal component of the displacement vector. The vertical component of the translation vector. The horizontal component of the translation vector;

[0032] S24: Generate a point spread function based on the translation and rotation components of the rigid body, and restore a clear image through Richardson-Lucy iterative deconvolution:

[0033] ;

[0034] in, For the image reconstructed in the nth iteration, For the image reconstructed in the (n+1)th iteration, The image is a blurry image that was actually captured. Let be the point spread function. Indicates pixel-by-pixel multiplication. This represents a convolution with a 180° rotation, and the number of iterations is less than or equal to 5. The termination condition is... .

[0035] Preferably, an image processing module is also included for performing:

[0036] S31: Perform nonlocal mean denoising on the visible light image output by the dynamic dual-mode imaging module, and perform anisotropic diffusion filtering on the infrared image. Finally, dynamically calculate the background model based on the time sliding window and subtract background interference by difference.

[0037] S32: A pre-trained U-Net neural network is used to segment the particle mask in the visible light image, and a temperature mask is adaptively generated based on the statistical characteristics of the infrared temperature distribution. Finally, the visible light and infrared masks are fused to generate an accurate target region.

[0038] S33: The particle displacement field is calculated using the optical flow method, a motion blur model is constructed based on the displacement field, and finally an iterative deconvolution algorithm is applied to reconstruct a high-resolution image;

[0039] S34: Input a continuous sequence of multiple time-series images, reconstruct a particle voxel space model through a three-dimensional convolutional neural network, and finally use an isosurface extraction algorithm to generate a three-dimensional morphology represented by a triangular mesh;

[0040] S35: Calculate the quantile parameters of particle size distribution, analyze surface sphericity and profile regularity, quantify the root mean square deviation of surface roughness, statistically analyze the area ratio of satellite powder adhesion, and measure the maximum temperature gradient value.

[0041] Preferably, the lateral wind field auxiliary detection unit includes the following components during operation:

[0042] S41: An adjustable lateral airflow is applied through an array of pneumatic nozzles orthogonal to the powder falling direction. The airflow velocity is adjusted by feedback from a mass flow controller and satisfies the following:

[0043] ;

[0044] in, It is a crosswind. This is the drag coefficient. The airflow density is the crosswind force. The projected area of ​​the particle. This refers to the lateral wind speed;

[0045] S42: Employs a high sampling rate laser displacement sensor array to measure the lateral displacement and time of particles under wind force in real time and record the trajectory equation;

[0046] S43: Deriving particle density based on Newton's second law:

[0047] ;

[0048] in, For the derived particle density, d is the particle diameter. Let be the lateral acceleration of the particle. This is the lateral displacement. The time when the displacement occurs;

[0049] S44: Input the derived particle density into the XRF composition-density relationship model for verification. The principle formula of the XRF composition-density relationship model is as follows:

[0050] ;

[0051] in, This represents the theoretical density of the particles. The number of alloying elements. Let i be the mass fraction of the i-th alloying element. Let be the density of the pure substance of the i-th alloying element. This is for porosity density correction.

[0052] As a preferred embodiment, the multi-dimensional feature fusion analysis module specifically includes:

[0053] A21: 3D-CNN spatiotemporal modeling unit, used to input continuous temporal image sequences and output particle 3D topology reconstruction results;

[0054] A22: Physical prior constraint module, used to embed the gas-atomized Navier-Stokes fluid dynamics equations as a loss function into the training of 3D-CNN spatiotemporal modeling units.

[0055] Preferably, the 3D-CNN spatiotemporal modeling unit includes the following process during operation:

[0056] S51: Temporal input processing, inputting a continuous temporal image sequence, and labeling the particle center coordinates of each frame image;

[0057] S52: Spatial alignment calibration, performs affine transformation on each frame of image based on the particle center coordinates to eliminate translation and rotation jitter;

[0058] S53: Four-dimensional tensor construction, stacking time series images into an H×W×T×1-dimensional tensor, where H and W are the height and width of the image, and performing normalization processing;

[0059] S54: Three-dimensional convolution kernel operation, extracting spatiotemporal features through cascaded 3D convolutional layers. The first layer consists of 32 5×5×3 convolutional kernels activated by ReLU, and the second layer consists of 64 3×3×3 convolutional kernels connected by residual connections.

[0060] S55: 3D topology reconstruction, outputting a 64×64×64 voxel mesh via a transposed convolutional layer Conv3DTrans;

[0061] S55: Surface extraction, using the moving cube algorithm to extract triangular meshes from a 64×64×64 voxel mesh at the isosurface = 0.5.

[0062] As a preferred option, the physical prior constraint module operates as follows:

[0063] S61: Fluid field physics modeling, which defines the fluid dynamics governing equations of the gas atomization process as prior knowledge for network training. These equations contain mathematical constraints between melt velocity field, pressure field, viscosity and gravity terms.

[0064] S62: Dynamic residual injection, which calculates the deviation between the actual physical field at each voxel location and the theoretical prediction of the fluid dynamics equation in the three-dimensional voxel mesh space output by 3D-CNN.

[0065] S63: A dual-objective optimization mechanism is constructed to build a joint loss function that integrates data error and physical deviation. Data error reflects the accuracy of image reconstruction, while physical deviation quantifies the degree to which the fluid equation is satisfied. The two are balanced by weighting coefficients.

[0066] S64: Intelligent constraint reinforcement, which predicts errors based on key quality indicators of the validation set and dynamically increases or decreases the weight coefficients of physical constraint terms.

[0067] S65: Backpropagation of physical perception. During the weight update process, the gradient of the physical deviation term is injected into the neural network through automatic differentiation, forcing the network parameters to satisfy the laws of fluid mechanics.

[0068] As a preferred embodiment, the multi-dimensional feature fusion analysis module, when in operation, specifically includes:

[0069] S71: Simultaneously receive real-time data streams from three independent sources, including the particle surface morphology feature set output by the visible light imaging module, the temperature field matrix and its gradient distribution output by the infrared thermal imager, and the dynamic parameter set output by the transverse wind field auxiliary detection unit.

[0070] S72: Based on a unified timestamp and spatial coordinates, the three types of data are aligned on a 4D spatiotemporal grid to eliminate transmission delay errors;

[0071] S73: Processed via a cascaded fusion engine, including:

[0072] A. First level: Physical association fusion, using decision tree rule base to establish logical mapping between surface morphology and temperature;

[0073] B. Second stage: Neural network feature fusion, inputting the three-modal feature vectors to a fully connected network, outputting the defect classification probability:

[0074] ;

[0075] in, This is the defect probability vector. This is the weight matrix. For morphological feature vectors, This is the temperature gradient feature vector. For dynamic eigenvectors, It is the bias vector;

[0076] C. Third level: Physical field inversion optimization, substituting the network output into the residual terms of the Navier-Stokes equation to correct misjudgments of fluid discontinuities;

[0077] S74: Based on the fusion results, four types of quality parameters are generated, including component purity parameters, structural integrity factors, sphericity optimization margin, and satellite powder risk level.

[0078] Preferably, the system also includes an online calibration module, which is used to spray standardized spherical particles into the powder stream at preset cycles and dynamically correct system parameters based on the measurement error of the standard particles. Specifically, it includes:

[0079] S81: Standardized spherical particles are sprayed into the main powder channel at a preset cycle through a high-pressure pneumatic nozzle array. The standardized particles have physical properties that meet the requirements of having a diameter of 50±1μm and being made of aluminum nitride or zirconium oxide.

[0080] S82: The dynamic dual-mode imaging module captures motion images of standard particles, and the calibration particles are screened out from the production powder based on the pre-stored three-dimensional feature template matching algorithm.

[0081] S83: Calculate the system measurement error for the identified calibration particles.

[0082] ;

[0083] in, This represents the three-dimensional error vector of the system measurement error. To account for particle size measurement error, For sphericity measurement error, For surface roughness error, The actual particle size value measured by the system. The sphericity measured by the system is the actual value. The actual surface roughness measured by the system. To provide a reference roughness value for calibrating the particles;

[0084] S84: Adjusts imaging parameters, measurement algorithms, and control thresholds in real time based on system measurement errors. Specifically:

[0085] A. Imaging parameters: When At that time, adjust the lens focus and lighting intensity;

[0086] B. Measurement algorithm, when At the same time, optimize the confidence threshold of the morphological segmentation model;

[0087] C. Control threshold, when Update the surface defect alarm trigger threshold at that time;

[0088] S85: Establish an error-parameter correction mapping database. When the same type of error occurs repeatedly, the calibration cycle will be shortened and the compensation range will be enhanced.

[0089] A multi-dimensional online analysis method for the sphericity of atomized powder includes the following steps:

[0090] S91: Simultaneously acquires the surface morphology and temperature field distribution of powder particles through visible light and infrared dual-mode imaging, dynamically adjusts the frame rate based on the particle movement speed, and uses optical flow method to correct motion blur.

[0091] S92: Apply orthogonal transverse wind force, combine with laser displacement sensor to measure particle deflection trajectory, invert particle density and verify internal defects;

[0092] S93: It integrates morphology, temperature and dynamic data, reconstructs the three-dimensional topology through 3D-CNN, and embeds fluid dynamics equations to constrain the training of AI models;

[0093] S94: Periodically spray standard particles for online calibration, and dynamically correct system parameters based on measurement errors.

[0094] The beneficial effects of this invention are:

[0095] 1. Compared with traditional density detection, which relies on destructive sampling and offline measurement, resulting in low efficiency and inability to cover all particles in the production line, this invention uses orthogonal wind field trajectory inversion technology to simultaneously output density values ​​and internal defect markers through non-contact dynamic measurement, achieving 100% non-destructive online full inspection of the production line, breaking through the sampling limitations and time bottlenecks of traditional methods.

[0096] 2. Compared with existing methods that are limited by surface morphology analysis and have insufficient recognition rate for deep defects such as internal pores and component segregation, this invention integrates morphology, temperature field and dynamic data, combined with physical constraint three-dimensional reconstruction technology, to penetrate the particle surface and realize visualization of internal structure, thus improving the ability to detect hidden defects to an industrial-level leading level.

[0097] 3. Compared with existing technologies, which suffer from blurred high-speed particle imaging due to fixed frame rate and exposure parameters, severely limiting the accuracy of sphericity analysis, this invention completely eliminates motion artifacts by using dynamic frame rate adjustment and microsecond-level flash freeze technology, combined with a sub-pixel displacement field reconstruction algorithm. This enables precise capture of the millimeter-level morphology of high-speed moving particles, providing a distortion-free data foundation for the quality assessment of atomized powder.

[0098] 4. Compared to existing systems that require periodic shutdowns for calibration, and whose fixed calibration cycles cannot suppress gradual drift, this invention uses online feedback from standard particles and error trend prediction to dynamically adjust optical parameters and algorithm thresholds, forming a self-evolving "measurement-calibration-optimization" mechanism to ensure long-term maintenance-free, high-precision operation of the system.

[0099] 5. Existing technologies rely on single-frame static analysis, which cannot capture the gradual change process of particle morphology, such as satellite powder adhesion and oxide layer growth. This invention analyzes the dynamic evolution law of particle surface topology through high-speed continuous frame time-series tracking, realizes early warning of gradual defects, and advances the timeliness of process degradation identification by 200%, breaking through the lag bottleneck of traditional quality control.

[0100] 6. Compared with traditional two-dimensional image analysis, which cannot reconstruct the internal structure of particles, resulting in a missed detection rate of over 40% for hidden defects such as pores and cracks; this invention uses 3D-CNN spatiotemporal modeling and reconstructs three-dimensional topology through voxelization of multiple frames of images to achieve visualization of submicron-level internal structures, thereby increasing the accuracy of hidden detection by a hundredfold.

[0101] 7. Compared to pure data-driven AI models, which are prone to failure when training data is insufficient and whose prediction results lack physical interpretability, this solution embeds the gas atomization Navier-Stokes equation as a physical loss function into the neural network training, forcing the model output to conform to the laws of fluid dynamics. It still maintains a prediction accuracy of over 95% with a small number of samples, and simultaneously improves the robustness and interpretability of the algorithm. Attached Figure Description

[0102] Figure 1 The diagram shown is a schematic representation of the structure of the online multi-dimensional analysis system for the sphericity of atomized powder according to the present invention.

[0103] Figure 2 The diagram shown is a flowchart of the online multi-dimensional analysis method for the sphericity of atomized powder according to the present invention. Detailed Implementation

[0104] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0105] Please see Figures 1-2 The present invention provides an embodiment: a multi-dimensional online analysis system for the sphericity of gas-atomized powder, comprising:

[0106] The dynamic dual-mode imaging module includes a visible light industrial camera and an infrared thermal imager, used to simultaneously acquire surface morphology images and temperature field distribution of atomized powder spheres.

[0107] Among them, the visible light industrial camera uses:

[0108] Telecentric lens combined with backlight / front light composite illumination;

[0109] High-frequency flashlights with pulse widths ≤1μs freeze particle motion;

[0110] An air curtain dustproof device protects optical components.

[0111] The dynamic dual-mode imaging module employs a high-speed adaptive imaging method, dynamically adjusting the frame rate of the visible light industrial camera based on the velocity of the atomized powder particles. The specific process is as follows:

[0112] The velocity of particles in the powder flow is collected in real time using a laser velocity sensor.

[0113] The target frame rate is calculated based on the movement speed of the atomized powder spheres using the frame rate calculation formula, which is:

[0114] ;

[0115] in, The velocity of the atomized powder particles. The minimum target particle size for atomized powder spheres. The average particle size of the atomized powder spheres. This is the fluctuation compensation coefficient. Target frame rate;

[0116] The frame rates of the visible light industrial camera and the infrared thermal imager are dynamically adjusted to the target frame rate, and the exposure time of the pulse flash is triggered synchronously. The pulse flash exposure time... The calculation formula is:

[0117] ;

[0118] Determine if a single particle has been continuously captured for at least 3 frames; if not, trigger secondary frame rate compensation.

[0119] ;

[0120] in, The frame rate after secondary compensation. The number of missing particles in the expected number of capture frames is counted in real time by the image matching algorithm.

[0121] Among them, the dynamic dual-mode imaging module adopts a multi-point sampling array, and the imaging points composed of the visible light camera and the infrared thermal imager are set to four or more groups, and the multiple imaging points are distributed along the transmission path of the atomized powder ball particles.

[0122] The dynamic dual-mode imaging module employs optical flow motion correction to eliminate the impact of motion blur on image analysis. The specific process is as follows:

[0123] Extract SIFT feature point sets from particle surfaces from continuously captured frame sequences;

[0124] Based on the improved Lucas-Kanade optical flow method, the displacement vector field of characteristic points is calculated. The principle formula is as follows:

[0125] ;

[0126] in, For the displacement vector field to be solved, For granular regions, To smooth the constraint coefficients, For image spatial gradient, The image's temporal partial derivative;

[0127] The displacement field is decomposed into rigid body translational and rotational components:

[0128] , ;

[0129] in, For rigid body translation components, This represents the number of effective feature points on the particle surface. The horizontal coordinates of the feature point in the image coordinate system. The vertical coordinates of the feature point in the image coordinate system. For point Displacement vector at that point For rotational components, It is a median function. This is the perpendicular component of the displacement vector. The horizontal component of the displacement vector. The vertical component of the translation vector. The horizontal component of the translation vector;

[0130] A point spread function is generated based on the translation and rotation components of the rigid body, and a clear image is restored through Richardson-Lucy iterative deconvolution.

[0131] ;

[0132] in, For the image reconstructed in the nth iteration, For the image reconstructed in the (n+1)th iteration, The image is a blurry image that was actually captured. Let be the point spread function. Indicates pixel-by-pixel multiplication. This represents a convolution with a 180° rotation, and the number of iterations is less than or equal to 5. The termination condition is... .

[0133] The steps used to extract the SIFT feature point set of the particle surface from the continuously captured frame sequence are as follows:

[0134] A Gaussian pyramid is constructed for the input image, and a multi-scale image sequence is generated by Gaussian convolution kernels of different scales. The pyramid has no less than 4 layers, and the scale parameter increases geometrically.

[0135] The Gaussian difference space is calculated by subtracting images of adjacent scales, and local extrema are searched in three-dimensional space, with each point compared with its 26 neighbors.

[0136] The sub-pixel positions of extreme points are corrected by fitting a quadratic function, and low-contrast points and edge response points are eliminated.

[0137] Calculate the gradient magnitude and direction in the neighborhood of key points, construct a direction histogram, and take the highest peak in the histogram whose peak exceeds a set proportion as the main direction.

[0138] Divide the region into 16×16 areas centered on key points, and further divide them into 4×4 sub-blocks. Calculate the 8-directional gradient histogram in each sub-block and stitch them together to form a 128-dimensional feature vector. Normalize the feature vector and truncate excessively large values.

[0139] Only retain points whose gradient magnitude exceeds a set proportion of the global maximum value; filter out non-target area points using a pre-segmented particle mask.

[0140] In summary, the dynamic dual-mode imaging module, when in operation, specifically includes:

[0141] A high-precision synchronous pulse signal is generated by an FPGA hardware timer, which simultaneously triggers the global shutter exposure of the visible light camera, the frame integration period of the infrared thermal imager, and the flash window of the dual-band pulse light source.

[0142] An incident light is separated into a visible light channel and an infrared channel using a beam splitter prism. The visible light is filtered by a 500nm high-pass filter and then the surface morphology is captured by a 2448×2048 resolution CMOS sensor. The infrared light is focused by a germanium lens onto a 640×512 resolution HgCdTe detector to record the temperature field.

[0143] Real-time geometric correction is performed on the dual-mode image based on pre-calibrated lens distortion parameters (radial and tangential coefficients) to eliminate optical distortion;

[0144] SIFT feature points from visible light images and phase-consistent feature points from infrared images are extracted, and pixel-level alignment is achieved through an affine transformation matrix.

[0145] Generate spatiotemporally aligned dual-mode data packets and transmit them to the image processing module.

[0146] In this embodiment, the dynamic dual-mode imaging module works in conjunction with a visible light industrial camera and an infrared thermal imager to simultaneously acquire the surface morphology and temperature field of atomized powder particles. The visible light camera uses a telecentric lens combined with backlight / frontlight composite illumination, supplemented by a high-frequency flash with a pulse width ≤1μs to freeze high-speed moving particles, and is equipped with an air curtain dustproof device to isolate metal dust; the infrared channel uses a germanium lens to accurately focus the thermal radiation signal. The module adopts a high-speed adaptive imaging method: based on a laser velocimetry sensor to acquire the particle movement speed in real time, the frame rate is dynamically adjusted (range 500-10000 Hz) to match particles of different sizes (e.g., 15-150μm), and a secondary compensation mechanism is triggered when the number of consecutive frames captured for a single particle is insufficient. Four or more imaging points are distributed along the powder transmission path to ensure multi-angle coverage. To address the motion blur problem, an optical flow method is used for correction: multi-scale SIFT feature points on the particle surface are extracted, the rigid body translation and rotation components are solved by an improved Lucas-Kanade algorithm, and a clear image is reconstructed by combining Richardson-Lucy deconvolution. All processes are synchronously controlled by FPGA hardware, and the final output is a time-space aligned dual-mode data packet (visible light image + infrared image + timestamp), with a transmission delay of less than 2ms.

[0147] Technical example: For 316L stainless steel powder (particle size 45μm), at a movement speed of 8m / s, the system automatically increases the frame rate to 8500 Hz, combined with 1μs flash freeze image, and after optical flow correction, the image sharpness is improved by 300% (MTF50 increases from 15lp / mm to 45lp / mm).

[0148] Technical Benefits: Adaptive High-Speed ​​Imaging: Frame rate dynamically matches particle velocity, ensuring 100% clear capture of particles larger than 15μm; Multimodal Precise Alignment: Spatial error between visible light and infrared images is less than 0.5 pixels, and time synchronization jitter is less than 1μs; Industrial-Grade Reliability: Triple dustproof design (air curtain isolation + nano-coating + electrostatic adsorption) adapts to... Dust concentration environment; Real-time guarantee: The entire process processing delay is less than 2ms, and it supports continuous analysis of 500Hz high-speed production lines.

[0149] The image processing module is used to perform:

[0150] Nonlocal mean denoising is performed on the visible light image output by the dynamic dual-mode imaging module, and anisotropic diffusion filtering is performed on the infrared image. Finally, the background model is dynamically calculated based on the time sliding window and the background interference is subtracted by difference.

[0151] A pre-trained U-Net neural network is used to segment the particle mask in the visible light image. A temperature mask is adaptively generated based on the statistical characteristics of the infrared temperature distribution. Finally, the visible light and infrared masks are fused to generate an accurate target region.

[0152] The particle displacement field is calculated using the optical flow method, a motion blur model is constructed based on the displacement field, and finally an iterative deconvolution algorithm is applied to reconstruct a high-resolution image.

[0153] Input a sequence of consecutive time-series images, reconstruct a particle voxel space model through a three-dimensional convolutional neural network, and finally use an isosurface extraction algorithm to generate a three-dimensional morphology represented by a triangular mesh.

[0154] Calculate the quantile parameters of particle size distribution, analyze surface sphericity and profile regularity, quantify the root mean square deviation of surface roughness, statistically analyze the area ratio of satellite powder adhesion, and measure the maximum temperature gradient value.

[0155] In this embodiment, the image processing module performs multi-level collaborative processing on the visible light and infrared images output by the dynamic dual-mode imaging module: First, visible light texture details are preserved through non-local mean denoising, while anisotropic diffusion filtering is used to suppress infrared thermal noise, and environmental interference is subtracted by dynamic background modeling; then, a pre-trained U-Net network is used to segment the particle morphology mask, and a thermal region mask is generated based on the statistical characteristics of infrared temperature distribution, and the target region is accurately locked through logical fusion; to address the motion blur problem, a blur model is constructed based on the displacement field calculated by the optical flow method, and an iterative deconvolution algorithm is applied to reconstruct the high-definition image; multiple consecutive frames of temporal images are input into the 3D-CNN network to reconstruct the three-dimensional voxel model of the particles, and the spatial morphology represented by the output triangular mesh is extracted through isosurface extraction; finally, multi-dimensional quality parameters such as particle size quantile (D10 / D50 / D90), sphericity, surface roughness (Ra), satellite powder coverage, and maximum temperature gradient are calculated.

[0156] Technical example: For cobalt-chromium alloy powder (particle size 80μm), the PSNR of the image after motion blur correction was improved from 24dB to 38dB, the internal pores (diameter 5.3μm) were detected by 3D reconstruction, and the surface roughness analysis accuracy reached 0.05μm.

[0157] Technical benefits: Improved multimodal fusion accuracy: Collaborative segmentation using topography and temperature masks reduces target area positioning error to less than 0.5 pixels; Dynamic blur correction: Motion reconstruction algorithm reduces the size measurement error of high-speed particles from ±9.1μm to ±0.8μm; Full-dimensional quality quantification: Synchronous output of 8 key parameters (such as sphericity accuracy ±0.5%), supporting real-time process optimization; Industrial real-time performance: Single particle processing time is less than 3ms, meeting the requirements for continuous analysis on 500Hz high-speed production lines.

[0158] The lateral wind field auxiliary detection unit is used to apply controllable lateral wind force through airflow nozzles orthogonal to the powder falling direction, and to measure the deflection trajectory of atomized powder particles in conjunction with a laser displacement sensor. Specifically, it includes:

[0159] An adjustable lateral airflow is applied through an array of pneumatic nozzles orthogonal to the powder's falling direction. The airflow velocity is regulated by feedback from a mass flow controller, and the following conditions are met:

[0160] ;

[0161] in, It is a crosswind. This is the drag coefficient. The airflow density is the crosswind force. The projected area of ​​the particle. This refers to the lateral wind speed;

[0162] A high-sampling-rate laser displacement sensor array is used to measure the lateral displacement and time of particles under wind force in real time, and record the trajectory equation:

[0163] ;in, Let be the lateral displacement of the particle at time t. t represents the initial lateral position of the particle, m represents the particle mass, and t represents the time the particle is subjected to wind force.

[0164] Derive particle density based on Newton's second law:

[0165] ;

[0166] in, For the derived particle density, d is the particle diameter. Let be the lateral acceleration of the particle. This is the lateral displacement. The time when the displacement occurs;

[0167] The derived particle density is input into the XRF composition-density relationship model for verification. The principle formula of the XRF composition-density relationship model is as follows:

[0168] ;

[0169] in, This represents the theoretical density of the particles. The number of alloying elements. Let i be the mass fraction of the i-th alloying element. Let be the density of the pure substance of the i-th alloying element. This is for porosity density correction.

[0170] Furthermore, during density inversion, density values ​​are output in real time and correlated with particle morphology data. When an anomaly is detected, the three-dimensional morphology data of the corresponding particle is automatically marked for re-inspection.

[0171] The lateral wind field auxiliary detection unit employs a multi-particle trajectory separation algorithm, specifically:

[0172] Firstly, based on the DBSCAN density clustering algorithm (neighborhood radius) 3) The original trajectory point cloud collected by the laser displacement sensor is segmented to separate spatially overlapping particle trajectories; then, the continuous motion of the same particle is matched by the first derivative of the displacement-time curve. If the slope deviation of adjacent time windows is less than 0.8 and the rate of change of acceleration is less than 0.8, the minimum number of core points is 3). If the trajectory is the same as the trajectory of the same particle, the complete trajectory sequence of the independent particles and their corresponding accelerations will be output.

[0173] The lateral wind field auxiliary detection unit employs a material adaptive mechanism. Specifically, a pre-set drag coefficient database covers parameters for over 200 powder materials, including metals, ceramics, and polymers. When the measured drag coefficient deviates from the theoretical value of the current material by more than 15%, composition inversion is automatically triggered. A fully connected neural network is used to analyze particle density, lateral acceleration, and particle size data, outputting the probability distributions of metals, ceramics, polymers, and mixed materials.

[0174] ;in, Let b be the weight matrix and b be the bias vector.

[0175] The material type is determined by the highest probability category, the drag coefficient setting is dynamically optimized, and the atomization process control system is linked: if it is determined to be a mixed material, the melt stirring frequency is increased by 20-50%; if the probability of polymer exceeds 70%, the atomization temperature is reduced by 10-30℃, so as to achieve a real-time closed-loop response to changes in material type.

[0176] In this embodiment, the lateral wind field auxiliary detection unit applies controllable lateral wind force through an array of pneumatic nozzles orthogonal to the powder falling direction, combined with a high-precision laser displacement sensor to measure the particle deflection trajectory in real time. The wind speed is dynamically adjusted by a mass flow controller (0.5-5 m / s), and the particle density is derived based on the Newtonian mechanics model: the lateral acceleration is calculated by the trajectory displacement and time difference, and the density value is derived by combining the wind force parameters. This density is compared and verified with the theoretical value of the XRF composition model (supporting 200+ alloy materials). When the deviation is greater than 5%, it is marked as a porosity defect and associated with the three-dimensional morphology data for re-examination. For multi-particle overlapping scenarios, DBSCAN clustering is used to segment the trajectory (neighborhood radius 10μm), and the motion of the same particle is matched by the continuity of the slope of the displacement-time curve. When the measured drag coefficient deviates from the theoretical value of the material by more than 15%, the neural network component inversion is triggered (input density, acceleration, particle size), and the probability distribution of metal / ceramic / polymer is output. The parameters are dynamically optimized and linked to the atomization process: the mixed material increases the melt stirring frequency by 20-50%, and the polymer material reduces the atomization temperature by 10-30℃, so as to realize the material adaptive closed-loop control.

[0177] Technical Example: For titanium alloy powder (Ti-6Al-4V): A 3.2 m / s transverse wind was applied, and the displacement was measured to be 42 μm / 4.5 ms. The density was then inverted. Derivation and XRF theoretical values A deviation of 0.23% is considered acceptable.

[0178] Technical benefits: Non-destructive density testing: Replaces traditional destructive sampling, improving efficiency by 100 times; Multi-particle trajectory decoupling: Overlapping trajectory separation success rate is greater than 98%, ensuring full inspection on high-speed production lines; Intelligent material identification: Component misjudgment rate is less than 2%, and process parameter self-adjustment response is less than 10ms; Precise defect location: Pore detection sensitivity reaches 10μm (CT verified), and three-dimensional morphology is marked for re-inspection simultaneously.

[0179] The multi-dimensional feature fusion analysis module is used to perform multi-dimensional feature analysis on atomized powder particles based on AI algorithms that fuse morphology, temperature, and kinetic data. Specifically, it includes:

[0180] The 3D-CNN spatiotemporal modeling unit is used to take continuous temporal image sequences as input and output 3D topological reconstruction results of particles.

[0181] The physical prior constraint module is used to embed the gas-atomized Navier-Stokes fluid dynamics equations as a loss function into the training of 3D-CNN spatiotemporal modeling units.

[0182] The 3D-CNN spatiotemporal modeling unit, when in operation, specifically includes the following process:

[0183] Temporal input processing: Input a continuous temporal image sequence, and annotate the particle center coordinates in each frame of the image;

[0184] Spatial alignment calibration involves performing an affine transformation on each frame of the image based on the particle center coordinates to eliminate translation and rotation jitter.

[0185] Four-dimensional tensor construction: The temporal images are stacked into an H×W×T×1-dimensional tensor, where H and W are the height and width of the image, and normalization is performed.

[0186] Three-dimensional convolutional kernel operations are used to extract spatiotemporal features through cascaded 3D convolutional layers. The first layer consists of 32 5×5×3 convolutional kernels activated by ReLU, and the second layer consists of 64 3×3×3 convolutional kernels connected by residual connections.

[0187] Three-dimensional topology reconstruction, outputting a 64×64×64 voxel mesh via a transposed convolutional layer Conv3DTrans;

[0188] For surface extraction, the moving cube algorithm is applied to extract triangular meshes from the 64×64×64 voxel mesh at the isosurface = 0.5.

[0189] The physical prior constraint module operates as follows:

[0190] Fluid field physics modeling defines the fluid dynamics governing equations of the gas atomization process as prior knowledge for network training. These equations contain mathematical constraints between melt velocity field, pressure field, viscosity, and gravity terms.

[0191] Dynamic residual injection calculates the deviation between the actual physical field at each voxel location and the theoretical prediction of the fluid dynamics equation in the three-dimensional voxel mesh space output by 3D-CNN.

[0192] A dual-objective optimization mechanism is proposed, which constructs a joint loss function that integrates data error and physical deviation. Data error reflects the accuracy of image reconstruction, while physical deviation quantifies the degree to which the fluid equation is satisfied. The two are balanced by weighting coefficients.

[0193] Intelligent constraint reinforcement dynamically increases or decreases the weight coefficient of physical constraint terms based on the prediction error of key quality indicators of the validation set.

[0194] In the backpropagation of physical perception, during the weight update process, the gradient of the physical deviation term is injected into the neural network through automatic differentiation, forcing the network parameters to satisfy the laws of fluid mechanics.

[0195] In summary, the multi-dimensional feature fusion analysis module, when in operation, specifically includes:

[0196] Simultaneously receive real-time data streams from three independent sources, including the particle surface morphology feature set output by the visible light imaging module, the temperature field matrix and its gradient distribution output by the infrared thermal imager, and the dynamic parameter set output by the transverse wind field auxiliary detection unit;

[0197] Based on a unified timestamp and spatial coordinates, the three types of data are aligned on a 4D spatiotemporal grid to eliminate transmission delay errors;

[0198] Processed through a cascading fusion engine, including:

[0199] A. First level: Physical association fusion, using a decision tree rule base to establish a logical mapping between surface morphology and temperature, for example:

[0200] If the surface pit depth is greater than 3μm and the local temperature gradient is greater than 200℃ / mm: it is judged as an internal porosity defect (confidence level 92%).

[0201] B. Second stage: Neural network feature fusion, inputting the three-modal feature vectors to a fully connected network, outputting the defect classification probability:

[0202] ;

[0203] in, This is the defect probability vector. This is the weight matrix. For morphological feature vectors, This is the temperature gradient feature vector. For dynamic eigenvectors, It is the bias vector;

[0204] C. Third level: Physical field inversion optimization, substituting the network output into the residual terms of the Navier-Stokes equation to correct misjudgments of fluid discontinuities;

[0205] Based on the fusion results, four types of quality parameters are generated: component purity parameter, structural integrity factor, sphericity optimization margin, and satellite powder risk level. Specifically:

[0206] The purity index of components uses a density-temperature correlation model to infer the degree of elemental segregation;

[0207] The structural integrity factor is calculated by combining the synergistic effect of surface pits and internal pores;

[0208] The sphericity optimization margin recommends the atomization pressure adjustment amount based on the aerodynamic drag coefficient;

[0209] The risk level of satellite dust is based on the prediction of satellite dust adhesion tendency using temperature gradient.

[0210] Furthermore, after generating four types of quality parameters based on the fusion results—including component purity parameters, structural integrity factors, sphericity optimization margin, and satellite powder risk level—a closed-loop control output is also included. Specifically, the quality parameters are transmitted to the atomizing device's PLC for execution in real time.

[0211] When the purity index of the components is less than 0.93, increase the melt superheating temperature by 10-50℃;

[0212] When the structural integrity factor is greater than 7.5, reduce the atomized gas-liquid ratio by 0.2-0.8.

[0213] In this embodiment, the multi-dimensional feature fusion analysis module processes continuous temporal images through a 3D-CNN spatiotemporal modeling unit: based on particle center coordinates, it aligns multiple frames of images to eliminate jitter, constructs a four-dimensional spatiotemporal tensor, extracts features through cascaded 3D convolutional layers, and finally outputs a 64×64×64 voxel grid, generating a triangular mesh 3D shape through a moving cube algorithm. The physical prior constraint module embeds the fluid dynamics equations of the gas atomization process (including constraints such as velocity field, pressure field, and viscosity) into the training, calculates the residuals between the actual physical field and the theoretical values, constructs a dynamic weighted loss function in conjunction with data reconstruction errors, and forces the network parameters to conform to the laws of fluid mechanics through automatic differentiation and backpropagation. The module synchronously receives morphology, temperature field, and dynamic data. After alignment with a unified spatiotemporal grid, it is processed by a three-level fusion engine: decision tree rules associate morphology and temperature characteristics (e.g., marking pore defects when pit depth is greater than 3μm and temperature gradient is greater than 200℃ / mm), a fully connected network outputs the defect probability distribution, and the Navier-Stokes equation residual corrects misjudgments. Finally, it generates four types of parameters: component purity index (inversely estimating elemental segregation), structural integrity factor (quantifying the synergistic effect of surface and internal defects), sphericity optimization margin (recommending atomization pressure adjustment amount), and satellite powder risk level (predicting adhesion tendency). It also links the atomization equipment PLC in real time: when the component purity is less than 0.93, it increases the melt temperature by 10-50℃; when the structural integrity factor is greater than 7.5, it reduces the gas-liquid ratio by 0.2-0.8, realizing closed-loop process optimization.

[0214] Technical Example: For titanium alloy powder (Ti-6Al-4V), 3D-CNN reconstruction detected internal pores with a diameter of 8μm. The decision tree correlated surface pits (5.2μm) with local low-temperature regions (ΔT=120℃), triggering a pore defect alarm.

[0215] For N718 high-temperature alloy powder, the satellite powder risk level exceeded the threshold, and the atomization pressure was automatically adjusted by +0.6 bar, reducing the satellite powder coverage from 18% to 5%.

[0216] Technical Results: 3D Defect Vision: Internal pore detection rate of 98.5% (25% improvement over traditional methods), positioning accuracy ±1.2μm; Multi-physics Collaboration: Fluid dynamic constraints reduce training data requirements by 60%, and network prediction error is less than 8%; Intelligent Process Recommendation: Sphericity optimization margin guides parameter adjustment, reducing scrap rate by 37%; Real-time Closed-Loop Control: Response time from defect identification to process adjustment is less than 0.5 seconds, improving production line yield by 12%.

[0217] The online calibration module is used to spray standardized spherical particles into the powder stream at preset cycles and dynamically correct system parameters based on the measurement error of the standard particles. Specifically, it includes:

[0218] Standardized spherical particles are sprayed into the main powder channel at a preset cycle through a high-pressure pneumatic nozzle array. The standardized particles have physical properties that meet the requirements of having a diameter of 50±1μm and being made of aluminum nitride or zirconium oxide.

[0219] The motion images of standard particles are captured using a dynamic dual-mode imaging module, and calibration particles are screened out from the production powder based on a pre-stored three-dimensional feature template matching algorithm.

[0220] Calculate the system measurement error for the identified calibration particles:

[0221] ;

[0222] in, This represents the three-dimensional error vector of the system measurement error. To account for particle size measurement error, For sphericity measurement error, For surface roughness error, The actual particle size value measured by the system. The sphericity measured by the system is the actual value. The actual surface roughness measured by the system. To provide a reference roughness value for calibrating the particles;

[0223] Imaging parameters, measurement algorithms, and control thresholds are adjusted in real time based on system measurement errors. Specifically:

[0224] A. Imaging parameters: When At that time, adjust the lens focus and lighting intensity;

[0225] B. Measurement algorithm, when At the same time, optimize the confidence threshold of the morphological segmentation model;

[0226] C. Control threshold, when Update the surface defect alarm trigger threshold at that time;

[0227] Establish an error-parameter correction mapping database to automatically shorten the calibration cycle and enhance the compensation magnitude when the same type of error occurs repeatedly.

[0228] Specifically, when using a dynamic dual-mode imaging module to capture motion images of standard particles and then using a pre-stored three-dimensional feature template matching algorithm to sieve calibration particles from production powder, the process includes:

[0229] The standard particles were rotated 360° and scanned to acquire high-resolution images from multiple perspectives.

[0230] Extract SIFT feature point sets and perform principal component analysis to reduce dimensionality and generate a 128-dimensional feature template library, storing three-dimensional signatures of surface morphology, texture distribution and contour curvature;

[0231] The high-speed imaging module is triggered to capture the sequence of falling particles and performs background subtraction and non-local mean denoising in real time to enhance the contrast of particle edges.

[0232] Perform multi-level feature matching, specifically:

[0233] A. Initial screening: Using Hu moment to quickly match the particle projection profile, candidate particles with a sphericity greater than 0.95 are screened.

[0234] B. Fine screening: Extract SIFT feature points from candidate particles, perform PCA similarity calculation with the template library, and set the cosine similarity threshold to be greater than 0.92;

[0235] C. Validation: Perform RANSAC robust fitting on the matching point pairs to eliminate false matches;

[0236] The screened calibration particles are marked with spatiotemporal IDs and their motion trajectories are continuously tracked for more than 3 frames to verify that the displacement conforms to the law of free fall.

[0237] The system outputs the real-time position coordinates and morphology parameters of the calibration particles, and triggers a system self-test alarm when five consecutive particles fail to match.

[0238] In this embodiment, the online calibration module periodically sprays standardized spherical particles (50±1μm in diameter, made of aluminum nitride or zirconium oxide) into the powder stream using a high-pressure pneumatic nozzle. The module captures the motion images of these particles using a dynamic dual-mode imaging module. Based on a pre-stored three-dimensional feature template (extracted through 360° scanning using SIFT-PCA dimensionality reduction signature), calibration particles are screened from the production powder. First, candidate particles with a sphericity greater than 0.95 are initially screened using Hu moment profile matching. Then, PCA similarity calculation (threshold greater than 0.92) and RANSAC robust fitting are used for fine screening. Finally, the trajectory is verified to conform to the free-fall law. For the identified calibration particles, three-dimensional error vectors of particle size, sphericity, and roughness are calculated, and system parameters are corrected in real time—the lens focal length and illumination are adjusted when the particle size error is greater than 3%, the segmentation algorithm confidence threshold is optimized when the sphericity error is greater than 0.01, and the defect alarm threshold is updated when the roughness error is greater than 0.1μm. Establish an error-parameter mapping database. When the same type of error occurs repeatedly, the calibration cycle will be shortened and the compensation will be increased (e.g., if the error is consistently high, the cycle will be shortened by 50%).

[0239] Technical Example: In a 316L stainless steel powder production line, 5 standard particles are sprayed every 30 minutes with a recognition rate of 98.7%. When a particle size system deviation of +3.5% (due to lens thermal drift) is detected, automatic focusing compensation is performed, and the deviation is eventually reduced to 0.2%.

[0240] In the high-temperature environment of titanium powder, the roughness error exceeded the threshold three times in a row, triggering the calibration cycle to be shortened from 60 minutes to 25 minutes, the lighting intensity to be increased by 20%, and finally the error returned to the normal range.

[0241] Technical benefits: Intelligent error suppression: Real-time correction of system drift reduces long-term measurement fluctuations from ±5% to ±0.3%; Anti-interference particle screening: Three-dimensional template matching + RANSAC verification achieves a standard particle identification accuracy greater than 99.5% (dust concentration). Operating conditions); self-evolutionary calibration: error trend prediction triggers dynamic adjustment of the cycle, reducing maintenance costs by 60%; seamless closed loop: parameter adjustment delay is less than 10ms, ensuring continuous operation of the production line without shutdown.

[0242] A multi-dimensional online analysis method for the sphericity of atomized powder includes the following steps:

[0243] The surface morphology and temperature field distribution of powder particles are acquired simultaneously by visible light and infrared dual-mode imaging. The frame rate is dynamically adjusted based on the particle motion speed, and the motion blur is corrected by optical flow method.

[0244] By applying orthogonal transverse wind force and combining it with laser displacement sensor to measure particle deflection trajectory, particle density is inverted and internal defects are verified.

[0245] By integrating morphology, temperature, and dynamic data, a three-dimensional topological structure is reconstructed using 3D-CNN, and fluid dynamics equations are embedded to constrain the training of the AI ​​model.

[0246] The system is calibrated online by periodically spraying standard particles, and the system parameters are dynamically corrected based on the measurement error.

[0247] In this embodiment, the multi-dimensional online analysis method for the sphericity of atomized powder is implemented by simultaneously acquiring the surface morphology and temperature field distribution of particles through visible light and infrared dual-mode imaging. The frame rate (500-10000Hz) is dynamically adjusted based on real-time particle velocity, and motion blur is corrected using optical flow. Orthogonal transverse wind force is applied in combination with a laser displacement sensor to measure the particle deflection trajectory, invert the density, and verify internal pore defects (e.g., marking when the density deviation is greater than 5%). The morphology, temperature, and dynamic data are fused, and the three-dimensional topology is reconstructed using 3D-CNN and AI training is constrained by fluid dynamics equations to improve defect recognition accuracy (e.g., internal pore detection rate of 98.5%). Standard particles (50±1μm aluminum nitride spheres) are periodically sprayed for online calibration, and system parameters are dynamically corrected based on particle size, sphericity, and roughness errors (e.g., automatic focusing when particle size error is greater than 3%) to achieve self-optimizing closed-loop control.

[0248] Example 1

[0249] Scenario: An aero-engine blade manufacturer uses a gas atomization process to produce Ti-6Al-4V titanium alloy powder. Traditional manual sampling involves taking 1 kg of powder every 2 hours and sending it to the laboratory. Density testing using a hydrometer bottle takes 30 minutes, and CT scanning of internal pores takes 2 hours, resulting in production line delays and a 15% false negative rate. After deploying this system:

[0250] The dynamic dual-mode imaging module simultaneously captures the morphology and temperature field of 45μm particles at a frame rate of 9000fps. The optical flow method is used to correct and eliminate motion blur caused by a falling speed of 8m / s. The air curtain dustproof device can operate continuously for 48 hours in a high-temperature dust environment of 80℃ without pollution.

[0251] A transverse wind force of 3.5 m / s is applied to the transverse wind field unit. A laser displacement sensor measures the transverse displacement of a particle as 38 μm / 4.2 ms. The inverted density is then calculated. (0.2% deviation from the theoretical XRF value) is marked as acceptable;

[0252] The multi-dimensional fusion module detected another particle surface pit (5.2 μm deep) accompanied by a local low temperature region (ΔT=142℃). 3D-CNN reconstruction revealed an internal pore with a diameter of 8 μm. The physical constraint model correlated the deformation-heat transfer mechanism and determined it to be a solidification defect.

[0253] The online calibration module sprays 5 aluminum nitride standard balls (Ø50μm) every 30 minutes, and it was found that the system particle size measurement had a positive drift of 3.5%. After automatic focusing compensation, the error was reduced to 0.3%, and the calibration cycle was shortened to 20 minutes.

[0254] Closed-loop control triggers melt superheat of +35°C and atomization pressure of +0.7 bar for defective particles, reducing batch porosity from 0.8% to 0.1%, meeting the requirements of aviation standard AMS4999A.

[0255] Example 2

[0256] Scenario: A metal 3D printing service provider encountered a problem where excessive IN718 powder (greater than 15%) caused cracks in the printed parts. Traditional electron microscopy sampling requires stopping the machine for sampling, making it impossible to pinpoint the root cause of the process degradation.

[0257] The dual-mode imaging module's infrared channel detected micron-sized high-temperature particles (instantaneous temperature 1700℃) formed by molten droplet sputtering, and time-series analysis captured the gradual process of their attachment to the main particles.

[0258] The wind farm auxiliary unit detected that the density fluctuation of multiple particles exceeded ±5%, and the composition inversion module identified 3% ceramic impurities mixed in the raw materials (abnormal drag coefficient +23%), and the PLC was immediately activated to isolate the contaminated batch.

[0259] 3D-CNN reconstruction showed thermal stress cracks in the satellite powder adhesion area. The physical prior model, combined with the temperature gradient field (greater than 220℃ / mm) and aerodynamic drag changes, traced the source to unstable atomization pressure.

[0260] The self-calibration system triggers emergency calibration after switching materials, identifies roughness measurement drift (+0.15μm), and corrects it in real time by enhancing illumination and updating the segmentation threshold;

[0261] The system integrates multiple dimensions to output the risk level of satellite powder exceeding the threshold, automatically performs fine-tuning of atomization air pressure (+0.5 bar) and increase centrifugal disc speed (+200 rpm), reducing satellite powder coverage from 18% to 4% within 48 hours, meeting the ISO / ASTM 52900 standard.

[0262] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multi-dimensional online analysis system for the sphericity of atomized powder, characterized in that: include: The dynamic dual-mode imaging module includes a visible light industrial camera and an infrared thermal imager, used to simultaneously acquire surface morphology images and temperature field distribution of atomized powder spheres; The lateral wind field auxiliary detection unit is used to apply controllable lateral wind force through airflow nozzles orthogonal to the direction of powder fall, and combine it with a laser displacement sensor to measure the deflection trajectory of atomized powder balls. The multi-dimensional feature fusion analysis module is used to perform multi-dimensional feature analysis on gas-atomized powder particles using morphology, temperature, and kinetic data from a cascaded fusion engine. The cascaded fusion engine includes: The physical association fusion unit is used to establish a logical mapping between surface morphology and temperature using a decision tree rule base; The neural network feature fusion unit is used to input the topographic feature vector, temperature gradient feature vector, and dynamic feature vector into the fully connected network and output the defect classification probability. The physics field inversion optimization unit is used to substitute the network output into the residual terms of the Navier-Stokes equations to correct misjudgments of fluid discontinuities. Among them, the multi-dimensional feature fusion analysis module generates component purity parameters, structural integrity factors, sphericity optimization margins, and satellite powder risk levels based on the fusion results.

2. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 1, characterized in that: The dynamic dual-mode imaging module is configured as follows: A11: Employs a high-speed adaptive imaging method to dynamically adjust the frame rate of the visible light industrial camera based on the velocity of the atomized powder spheres. A12: It adopts a multi-point sampling array, and the imaging points composed of the visible light camera and the infrared thermal imager are set to four or more groups, and the multiple imaging points are distributed along the transmission path of the atomized powder ball particles. A13: Optical flow motion correction is used to eliminate the impact of motion blur on image analysis.

3. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 2, characterized in that: The specific process of the dynamic dual-mode imaging module in dynamically adjusting the frame rate of the visible light industrial camera based on the velocity of the atomized powder spheres is as follows: S11: Real-time acquisition of particle velocity in powder flow via laser velocity sensor; S12: Calculate the target frame rate based on the movement speed of the atomized powder spheres using the frame rate calculation formula; S13: Dynamically adjust the frame rate of the visible light industrial camera and the infrared thermal imager to the target frame rate, and simultaneously trigger the pulse flash exposure time; S14: Determine whether a single particle has been continuously captured for at least 3 frames. If not, trigger secondary frame rate compensation.

4. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 3, characterized in that: It also includes an image processing module for performing: S31: Perform nonlocal mean denoising on the visible light image output by the dynamic dual-mode imaging module, and perform anisotropic diffusion filtering on the infrared image. Finally, dynamically calculate the background model based on the time sliding window and subtract background interference by difference. S32: A pre-trained U-Net neural network is used to segment the particle mask in the visible light image, and a temperature mask is adaptively generated based on the statistical characteristics of the infrared temperature distribution. Finally, the visible light and infrared masks are fused to generate an accurate target region. S33: The particle displacement field is calculated using the optical flow method, a motion blur model is constructed based on the displacement field, and finally an iterative deconvolution algorithm is applied to reconstruct a high-resolution image; S34: Input a continuous sequence of multiple time-series images, reconstruct a particle voxel space model through a three-dimensional convolutional neural network, and finally use an isosurface extraction algorithm to generate a three-dimensional morphology represented by a triangular mesh; S35: Calculate the quantile parameters of particle size distribution, analyze surface sphericity and profile regularity, quantify the root mean square deviation of surface roughness, statistically analyze the area ratio of satellite powder adhesion, and measure the maximum temperature gradient value.

5. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 4, characterized in that: The lateral wind field auxiliary detection unit, when in operation, specifically includes: S41: An adjustable lateral airflow is applied through an array of pneumatic nozzles orthogonal to the direction of powder fall, and the airflow speed is adjusted by feedback from a mass flow controller. S42: Employs a high sampling rate laser displacement sensor array to measure the lateral displacement and time of particles under wind force in real time and record the trajectory equation; S43: Derive particle density based on Newton's second law; S44: Input the derived particle density into the XRF composition-density relationship model for verification.

6. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 5, characterized in that: The multi-dimensional feature fusion analysis module specifically includes: A21: 3D-CNN spatiotemporal modeling unit, used to input continuous temporal image sequences and output particle 3D topology reconstruction results; A22: Physical prior constraint module, used to embed the gas-atomized Navier-Stokes fluid dynamics equations as a loss function into the training of 3D-CNN spatiotemporal modeling units.

7. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 6, characterized in that: The spatiotemporal modeling unit of a 3D-CNN operates through the following process: S51: Temporal input processing, inputting a continuous temporal image sequence, and labeling the particle center coordinates of each frame image; S52: Spatial alignment calibration, performs affine transformation on each frame of image based on the particle center coordinates to eliminate translation and rotation jitter; S53: Four-dimensional tensor construction, stacking time series images into an H×W×T×1-dimensional tensor, where H and W are the height and width of the image, and performing normalization processing; S54: Three-dimensional convolution kernel operation, extracting spatiotemporal features through cascaded 3D convolutional layers; S55: 3D topology reconstruction, outputting a 64×64×64 voxel mesh via a transposed convolutional layer Conv3DTrans; S55: Surface extraction, using the moving cube algorithm to extract triangular meshes from a 64×64×64 voxel mesh at the isosurface = 0.

5.

8. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 7, characterized in that: The multi-dimensional feature fusion analysis module, when in operation, specifically includes: S71: Simultaneously receive the particle surface morphology feature set output by the visible light imaging module, the temperature field matrix and its gradient distribution output by the infrared thermal imager, and the dynamic parameter set output by the transverse wind field auxiliary detection unit; S72: Align the three types of data on a 4D spatiotemporal grid based on a unified timestamp and spatial coordinates; S73: Processed via a cascaded fusion engine; S74: Based on the fusion results, four types of quality parameters are generated, including component purity parameters, structural integrity factors, sphericity optimization margin, and satellite powder risk level.

9. The online multi-dimensional analysis system for the sphericity of atomized powder according to claim 8, characterized in that: It also includes an online calibration module, which is used to spray standardized spherical particles into the powder stream at preset cycles and dynamically correct system parameters based on the measurement error of the standard particles. Specifically, it includes: S81: Standardized spherical particles are sprayed into the main powder channel at a preset cycle through a high-pressure pneumatic nozzle array; S82: The dynamic dual-mode imaging module captures motion images of standard particles, and the calibration particles are screened out from the production powder based on the pre-stored three-dimensional feature template matching algorithm. S83: Calculate the system measurement error for the identified calibration particles; S84: Adjusts imaging parameters, measurement algorithms, and control thresholds in real time based on system measurement errors; S85: Establish an error-parameter correction mapping database. When the same type of error occurs repeatedly, the calibration cycle will be shortened and the compensation range will be enhanced.

10. A multi-dimensional online analysis method for the sphericity of atomized powder, applied to the multi-dimensional online analysis system for the sphericity of atomized powder as described in any one of claims 1 to 9, characterized in that: Includes the following steps: S91: Simultaneously acquires the surface morphology and temperature field distribution of powder particles through visible light and infrared dual-mode imaging, dynamically adjusts the frame rate based on the particle movement speed, and uses optical flow method to correct motion blur. S92: Apply orthogonal transverse wind force, combine with laser displacement sensor to measure particle deflection trajectory, invert particle density and verify internal defects; S93: Fuses topographic, temperature, and dynamic data, processed through a cascaded fusion engine, including: Data is aligned on a 4D spatiotemporal grid based on a unified timestamp and spatial coordinates; Perform physical correlation fusion, neural network feature fusion, and physical field inversion optimization; Generate component purity parameters, structural integrity factor, sphericity optimization margin, and satellite powder risk level; S94: Periodically spray standard particles for online calibration, and dynamically correct system parameters based on measurement errors.

Images